Differentiable External-Validation Technical Report
July 22, 2026 ยท View on GitHub
This report defines the public comparison and reproducibility package for the differentiable-programming lane. It is a technical report, not a promotion claim. Every row remains bounded by the committed claim ledger until external comparison rows, the claim ledger, and isolated benchmark artefacts all pass.
Claim Boundary
Current evidence is functional_non_isolated unless an artefact explicitly says
otherwise. The report does not claim production performance, quantum advantage,
hardware execution, provider execution, QPU execution, GPU execution, or
isolated_affinity benchmark status.
The public wording source is
data/differentiable_phase_qnode/public_claim_table_20260616.md. The reviewer
ledger is data/differentiable_phase_qnode/claim_ledger.md.
Reproducibility Package
| Artefact | Role | Boundary |
|---|---|---|
data/differentiable_phase_qnode/external_validation_artifact_bundle_20260616.json | SHA-256 manifest over committed differentiable evidence files. | Checksum provenance only. |
data/differentiable_phase_qnode/external_validation_environment_lock_20260616.json | Exact runtime, developer, CI, framework-overlay, and Enzyme-runner lockfile manifest. | Reviewer reproduction only. |
data/differentiable_phase_qnode/local_benchmark_20260616T0955Z/diff-qnode-external-comparison.json | Bounded JAX, PyTorch, PennyLane, Enzyme, TensorFlow, and Catalyst comparison rows, including the dedicated Catalyst compiler-workflow profile added during the 2026-07-04 schema refresh. | Functional non-isolated comparison evidence. |
data/differentiable_phase_qnode/identical_circuit_gradient_comparison_20260616.json | Same-circuit gradient comparison artefact for Qiskit and PennyLane routes. | Correctness comparison, not hardware execution. |
data/differentiable_phase_qnode/domain_benchmark_dataset_closure_20260616.json | Exact-answer domain dataset closure artefact. | Dataset validation, not production benchmark promotion. |
data/differentiable_phase_qnode/differentiable_architecture_map_20260627.json | Architecture and Rustification routing map over inventory rows and scorecard categories. | Routing evidence only, not Rust or performance promotion. |
data/differentiable_phase_qnode/differentiable_dependency_environment_map_20260627.json | Dependency and environment evidence map over lock profiles, SHA-bound toolchain versions, and local/cloud/provider/hardware/GTX 1060/ML350 route classifications. | Dependency and route provenance only; unlocked GPU and deferred execution rows remain hard gaps, not promotion. |
data/differentiable_phase_qnode/differentiable_isolated_benchmark_plan_20260627.json | Reserved-host batch plan over current non-isolated differentiable benchmark/evidence artefacts, including the compiler-promotion batch gate. | Planning evidence only, not benchmark execution or isolated_affinity promotion. |
data/differentiable_phase_qnode/provider_gradient_boundary_20260705.json | No-submit provider-gradient, hardware-policy, and provider-preparation boundary evidence. | Local callback and governance evidence only, not live provider, QPU, hardware-gradient result, isolated benchmark, or performance promotion. |
data/differentiable_phase_qnode/compiler_evidence_boundary_20260705.json | Native LLVM/JIT focused-test evidence plus current Enzyme/LLVM scalar, vector, matrix, loop execution, and Program AD alias-activity boundary. | Compiler boundary evidence only; not general compiler-AD, isolated benchmark, provider, hardware, GPU, or performance promotion. |
data/differentiable_phase_qnode/compiler_alias_activity_evidence_20260706.json | Program AD static alias-lattice evidence covering bounded view, list, scalar-rebinding, object-attribute, loop-carried, branch-control, and mutation-version alias activity. | Alias-activity compiler evidence only; not general compiler-AD, isolated benchmark, provider, hardware, GPU, or performance promotion. |
data/differentiable_phase_qnode/compiler_promotion_batch_20260706.json | Non-promotional compiler evidence batch that checksums the compiler boundary, alias-activity, native LLVM/JIT, native whole-program AD, Enzyme/MLIR maturity, and raw Enzyme/MLIR breadth evidence. | Batch assembly evidence only; still blocked by missing isolated compiler benchmark artifact IDs. |
data/differentiable_phase_qnode/enzyme_mlir_compiler_ad_breadth_artifact_20260706.json | Raw 11-case Enzyme/MLIR compiler-AD breadth artifact with passing scalar reverse, vector VJP, matrix VJP, loop, LLVM IR generation, and native Enzyme rows plus explicit hard gaps for missing scalar forward, vector JVP, matrix JVP, MLIR lowering runtime, and Enzyme/MLIR alias-activity raw cases. | Case-level breadth evidence only; non-promotional until isolated benchmark and derived breadth evidence pass. |
data/differentiable_phase_qnode/enzyme_mlir_maturity_audit_20260616.json | Enzyme/MLIR maturity audit with MLIR-runtime correctness evidence, native Enzyme scalar evidence, a linked raw 11-case compiler-AD breadth artifact, and missing derived breadth evidence recorded as hard gaps. Partial breadth captures are represented through complete artifacts with explicit case hard gaps before promotion. | Hard-gap evidence, not Enzyme parity promotion. |
Provider Comparison Status
| Provider family | Current SCPN evidence | Still blocked before promotion |
|---|---|---|
| JAX | CPU overlay comparison rows and bounded bridge evidence exist for declared routes. | Mature native arbitrary Phase-QNode lowering, broad transform algebra, and isolated benchmark artefacts. |
| PyTorch | Bounded parameter-shift, custom autograd, torch.func, torch.compile, module/layer wrapper, and live-overlay evidence exist for declared routes. | Promotion-grade isolated benchmark artefacts and arbitrary simulator/compiler lowering. |
| TensorFlow | Bounded GradientTape, tf.function, XLA, and Keras-layer evidence exist for declared routes. | Arbitrary Phase-QNode lowering and isolated benchmark artefacts. |
| PennyLane | Export/import and identical-circuit comparison evidence exists for bounded local circuits. | Plugin/provider hardware routes and promotion-grade isolated benchmark evidence. |
| Qiskit | Shifted-circuit generation, local Statevector, finite-shot surrogate, and no-submit provider-gradient/preparation boundary evidence exist. | Live execution, raw-count replay, calibration/statevector comparison, and hardware-gradient artefacts. |
| Enzyme/LLVM/MLIR | MLIR-runtime correctness evidence is attached; the local enzyme, opt, mlir-opt, and clang command/version snapshot is recorded; bounded native LLVM/JIT focused tests pass under the prescribed selector; native LLVM/JIT crash-safety test IDs are attached; rollback/fallback policy is attached; current LLVMEnzyme execution covers scalar, vector, matrix, and loop kernels with zero gradient error; the raw 11-case Enzyme/MLIR breadth artifact is attached with explicit hard-gap rows; bounded Program AD alias-activity evidence covers static alias-lattice complete and fail-closed cases; the compiler promotion batch is assembled as non-promotional checksum evidence. | Isolated compiler benchmark IDs, derived arbitrary Enzyme/program AD breadth evidence, and the Enzyme-JAX external-comparison runtime gap. |
| Catalyst | The external-comparison artefact now emits a Catalyst hard-gap row with qjit/MLIR/QIR workflow scope, compiled differentiation scope, control-flow gaps, finite-shot limitations, and unsupported provider routes. | Configured Catalyst runner success, arbitrary compiled quantum-classical workflow parity, finite-shot jobs, provider submission, hardware execution, and isolated benchmark IDs. |
Reviewer Procedure
- Validate the public claim table before quoting differentiable capability language.
- Validate the environment lock before replaying framework comparisons.
- Validate the artefact bundle before using any evidence file in review.
- Treat local benchmark rows as regression and functional comparison evidence
unless the artefact itself reports
isolated_affinity.
The current package is therefore suitable for reviewer reproduction and gap triage. It is not yet suitable for performance promotion.